7.9.35. MNIST Image Classification

Handwritten digit recognition (0-9) on the MSPM0G5187 microcontroller using the classic LeNet-5 CNN.

Example folder: examples/MNIST_image_classification (GitHub)

7.9.35.1. Overview

Handwritten digit recognition is a basic task in document processing and automation. Traditional template matching struggles with handwriting variability, and edge AI allows recognition directly on a resource-constrained MCU without cloud connectivity.

Task: image classification (image_classification, target_module: vision), 10 classes
Device: MSPM0G5187 (config_image_classification_mnist.yaml, the only config in the folder)
Model: Lenet5 (about 60,000 parameters according to the readme)
Input: 28x28 grayscale images
Demo setup in the readme: the MCU talks to a host GUI over UART

7.9.35.2. The Dataset

The data is MNIST exported with torchvision as 28x28 grayscale PNG images in 10 class folders. The config downloads https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/mnist_classes.zip automatically. To recreate the dataset, run mnist_dataset_creation.py in the example folder; see readme_dataset_creation.md there. The layout is a classes/ folder with one sub-folder per digit.

7.9.35.3. Feature Extraction

The config uses the preset Mnist_Default. The preprocessing converts to grayscale (1 channel), resizes to 28x28, converts to a tensor and normalizes with mean 0.1307 and scale 0.3081.

7.9.35.4. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/MNIST_image_classification/config_image_classification_mnist.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\MNIST_image_classification\config_image_classification_mnist.yaml

7.9.35.5. Configuration

common:
  target_module: vision
  task_type: image_classification
  target_device: MSPM0G5187
dataset:
  dataset_name: mnist_image_classification
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/mnist_classes.zip
data_processing_feature_extraction:
  feature_extraction_name: Mnist_Default
training:
  model_name: Lenet5
  batch_size: 64
  training_epochs: 14
  num_gpus: 1
  quantization: 2
  learning_rate: 0.1
testing: {}
compilation: {}

quantization: 2 gives INT8 quantization for a smaller model. num_gpus: 1 expects a GPU; use 0 for CPU. The image models enabled in 1.5.0 are Lenet5, MobileNetV1_28k_NPU and MobileNetV2_58k_NPU. For the image task type see Image Classification.

7.9.35.6. Results

The readme quotes an accuracy of about 99%. The automated 1.5.0 config test run reached a best quantized epoch of 3 with a validation accuracy of 98.94% (F1 0.989, AUC ROC 0.998) and a test accuracy of 98.93% (AUC ROC 0.999).

7.9.35.7. Output Location

Results are written to ./data/projects/mnist_image_classification/run/<date-time>/Lenet5/. See Understanding the Run Output and Project Compilation Guide: Timeseries Classification Examples for deployment with the MSPM0 SDK (2.08.00 or later) and Code Composer Studio.